meta_test_head is None on demo

[open-mmlab]

meta_test_head is None on demo

[linyq17]

The meta testing of NegMargin will build and finetune a new head for each task. Currently, the demo does not support methods that require fine-tuning during testing since the inference codes do not introduce any code of training.

[linyq17]

If you only need to test a few tasks, it might work by saving all the pre-trained weights. As for the inference, I think it is possible to compute the weights of fc layer from extracted features mutually since it only trains a linear head.

[rlleshi]

What if we process support images for the same task as was during training and validation? In this case, a new head won't be needed because it has already been fine-tuned during the meta testing of the training phase. In other words, the head formed during meta-test in training could be directly utilized and the step of fine-tuning a new head skipped. Does this make sense?

Alternatively, do you have any other suggestions to do inference while serving these models (which are performing better for my task) in production?